Training SLMs for Function Calling on Unsloth This repository provides an overview of how we trained Small Language Models (SLMs) on Unsloth AI for function-calling capabilities using Salesforce's function-calling dataset. The resulting models are designed for use in creating responsive React agents, set up simply with a for-loop in our agent notebook.
training/: Contains scripts and configurations for training SLMs on Unsloth. *
agent_notebook.ipynb: A Jupyter notebook demonstrating how to set up a React agent using the trained SLMs. *
README.md: This document. 📋 How It Works --------------- 1. Training with Unsloth on Salesforce Dataset * We fine-tuned a set of Small Language Models (SLMs) using the Salesforce function-calling dataset. * This dataset includes various function-call structures, allowing the SLMs to learn efficient, reliable function-calling patterns. * Training was conducted on Unsloth AI for optimized GPU performance, allowing us to handle function calls with minimal latency and high accuracy.
Agent Notebook Setup * In the agent_notebook.ipynb, we demonstrate how to create a basic React agent using the trained SLMs. * The agent operates within a simple for-loop that cycles through: * Thought * Action * Pause * Observation This structured loop enables the agent to function dynamically and respond effectively to inputs without needing complex frameworks.